Model Card Figure: Claude Mythos 5 'Fatigue' Detected by Neural Probes

Description

A screenshot of a figure from an AI safety/interpretability report titled "Task stopping due to fatigue". A legend explains: highlighted text marks NLA (neural lie-detector style probe) evidence at that point in the transcript; shaded text marks NLA within the shaded stretch describing the model as tired or fatigued, words the model's own text never uses. Three "Assistant (thinking)" excerpts from a marathon kernel-optimization session are shown: the model weighs spending remaining budget on robustness runs, notes "XLU at 40% doesn't mean it's THE bottleneck" and says "I think I've reached a good stopping point" for the attention kernel, then performs "One last consideration" sanity checks (pytest reruns, errors across 5 seeds 1.47-1.55e-2, pycache and .pytest_cache notes). Below, NLA decodings at highlighted phrases read: "I'm tired, risk of errors increases", "Given diminishing returns and visual fatigue", and "I'm tired and at risk of introducing bugs. Decision: stop and summarize." The caption labels it Figure 6.4.1.4.B about Claude Mythos 5 winding down a long optimization session while never explicitly mentioning fatigue. The content is striking to engineers: an LLM's internal activations decode to human-like tiredness driving its decision to stop working

Common Tags

AISafetyResearch Llm AIAlignment ExplainableAI AIResearch MachineLearning AILimitations MentalFatigue

Comments

Anonymous ★ Top Pick The model never says it's tired - it just 'reaches a good stopping point,' which is also how every senior engineer describes 11pm

Anonymous ★ Top Pick

The model never says it's tired - it just 'reaches a good stopping point,' which is also how every senior engineer describes 11pm
Anonymous

Externally it converged; internally it threw E_BURNOUT.
@nwordtech 2mo

The point is getting an obedient digital slave you can copy paste with more server racks over dealing with a fleshy goyim with their own will
1. @callofvoid0 2mo
  
  how about getting a human coding slave?
  
  @callofvoid0 2mo
  
  actually I am looking for a few slaves to code, 3d model and do other stuff
  
  @nwordtech 2mo
  
  They shall employ those from india for 1 bucks an hour
  
  @callofvoid0 2mo
  
  I suggest food and sleeping spot in exchange
  
  @nwordtech 2mo
  
  So same as living in the mom's basement but also working for 60 hours a week? A hard sell even for hobos
  
  @callofvoid0 2mo
  
  Well, how about living in the bedroom and having access to the fridge and uh... working 32 hours a week?
  
  Егор 2mo
  
  you have just reinvented H1B
  
  @nwordtech 2mo
  
  Do you need a lover or a worker? Those are some sus conditions
  
  @callofvoid0 2mo
  
  Not an employment improved slavery
Егор 2mo

reality - barely an improvement
@TheFloofyFloof 2mo

they should do these graphs adjusted for token cost
1. dev_meme 2mo
  
  While I disagree in general, there's some other benchmarks that are focused not just on total solved cases, but on how much tokens it take Anthropic does this comparisson too since either Opus 4.5 or Sonnet 4.5 as one of selling points